> ## Documentation Index
> Fetch the complete documentation index at: https://koharu.rs/llms.txt
> Use this file to discover all available pages before exploring further.

# Connect a translation provider

> Use local GGUF models, hosted services, or a compatible model server.

The translation picker combines local and hosted backends. A selection records the provider, model ID, optional local quantization, and vision-input setting.

## Available connections

| Connection          | Providers                                                                |
| ------------------- | ------------------------------------------------------------------------ |
| Local inference     | GGUF models through llama.cpp                                            |
| Hosted models       | Atlas Cloud, OpenAI, Gemini, Claude, Grok, MiniMax, DeepSeek, OpenRouter |
| Compatible servers  | LM Studio and OpenAI-compatible endpoints                                |
| Machine translation | DeepL, Google Cloud Translation, Caiyun                                  |

Each provider owns its endpoint defaults, request mapping, and model discovery. Available models can change independently of these docs.

<Steps>
  <Step title="Save the connection">
    Open **Settings → Providers**, choose a provider, and fill in its fields. Hosted services
    require credentials. A local compatible server needs its base URL; whether it needs a secret
    depends on that server.
  </Step>

  <Step title="Select a model">
    Open the processing selector or **Settings → Translation**. Opening the model picker refreshes
    provider discovery. An unreachable provider may have no discovered models even while its saved
    settings remain intact.
  </Step>

  <Step title="Choose whether to include the page image">
    Enable **Vision input** only for a model that accepts images. Cataloged text-only models disable
    the switch. OpenAI-compatible endpoints are user-declared, so verify the endpoint and model
    support image messages.
  </Step>

  <Step title="Try one page">
    Check the source OCR, run translation on the current page, and review the response before
    processing a chapter.
  </Step>
</Steps>

## Credentials and privacy

Provider secrets use the operating system's secure credential service. Non-secret endpoint settings live in `~/.koharu/config.toml`.

Local models keep prompts and text on your machine after download. Hosted providers receive request content and apply their own retention terms. Local detection and OCR do not make a hosted translation request local.

<Warning>
  Keep keys out of project instructions, text layers, and configuration files. Save credentials
  through **Settings → Providers**.
</Warning>